Towards Nonlinear Multimaterial Topology Optimization Using Unsupervised Machine Learning and Metamodel-Based Optimization

This work introduces a multimaterial density-based topology optimization method suitable for nonlinear structural problems. The proposed method consists of three stages: continuous density distribution, clustering, and metamodel-based optimization. The initial continuous density distribution is generated following a synthesis strategy without penalization, e.g., the hybrid cellular automaton (HCA) method. In the clustering stage, unsupervised machine learning (e.g., K-means clustering) is used to optimally classify the continuous density distribution into a finite number of clusters based on their similarity. Finally, a metamodel (e.g., Kriging interpolation) is generated and iteratively updated following a global optimization algorithm (e.g., genetic algorithms) to ultimately converge to an optimal material distribution. The proposed methodology is demonstrated with the design of multimaterial stiff (minimum compliance) structures, compliant mechanisms, and a thin-walled S-rail structure for crashworthiness.